[Mesa-dev] NIR + softfp64 problem

[Gert Wollny]

Hello Matt, 

Am Mittwoch, den 16.01.2019, 10:17 -0800 schrieb Matt Turner:
> any idea how to quell this would be very welcome.
> 
> It's required to scalarize fp64 operations before this lowering code
> will work. It looks to me like it's trying to call __fadd64 with a
> dvec4 argument, when the arguments are actually scalar.
> 
> I think r600 is mostly vector-based? The soft-fp64 code probably
> isn't ideally suited for that. I'd attempt to call
> nir_lower_alu_to_scalar() before calling nir_lower_doubles(). That
> should at minimum tell you whether my hypothesis is correct.
Indeed, r600 I don't do  nir_lower_alu_to_scalar() because it is more
convenient for me, but with that the number of crashes goes down
significantly, the remaining crashes seem to be in my own code. 
> 
> From there, maybe we could pass an options bitfield to
> nir_lower_alu_to_scalar() to allow R600 to only lower fp64 operations
> and not scalarize everything. 
I think this is the option I'll purse, I've touched that routine before
before I disabled it completely. 

> Or, we could try to figure out how to
> add vectorized versions of the soft-fp64 routines that R600 could use
> directly.
Unless other drivers can take advantage of this I don't think that this
woule make much sense. To get the r600 nir backend performant I'd have
to write some scheduler for the backend anyway (and this is still a
long way to go). 

thanks for the pointers, 
Gert